| Summary: | 博士 === 國立東華大學 === 應用數學系 === 97 === Metabolic syndrome is becoming increasing common worldwide and known to be an important risk factor for cardiovascular diseases, which are among the ten leading causes of death in Taiwan. Five metabolic measurements, namely blood pressure, fasting glucose, serum triglyceride, high density lipoprotein, and waist circumference, are commonly used to define metabolic syndrome: If any three measurements exceed their normal ranges, he or she is considered having the metabolic syndrome. It is apparent that the syndrome has its unique characteristics in Taiwanese. To study of the prevalence of the syndrome, the above metabolic measurements are used to determine the presence of individual metabolic disorders in many family follow-up survey studies. The risk factors of individual metabolic disorders are yet to be discovered.
Since individuals within a family may not only live together but also share partly the same inheritance of diseases, familial longitudinal data of the five metabolic disorders are clustered and multivariate in nature. Each family can be viewed as a cluster for such data. In addition, both follow-up data of a subject and the cross-sectional data of a survey cycle can form subclusters of a family. The clustering levels of subject and survey cycle are thus nested within families. These two levels are, however, crossed in the sense that either is not nested within the other. In addition to the usual risk assessment, our goal is to explore which level of clustering induces largely the intra-family correlation for each disorder, and to identify factors that may influence the concurrent pairwise association among disorders.
Commonly used univariate regression models are often appropriate only for clustered data with multiple levels of nesting. To analyze familial longitudinal data of metabolic disorders that are of nonnesting structure, we propose a multivariate regression model that use (1) the marginal regression to conduct the risk assessment
for each metabolic disorder, and (2) the log-linear odds ratios regression to study the concurrent pairwise-association among disorders. To investigate the source of intra-family correlation, each disorder is further assumed to be the manifestation of a latent normal variable with unit marginal variance through a common threshold. The latent mean is determined by the aforementioned marginal regression. According to the clustering levels of the data at hand, the random component of the latent variable is decomposed into several random effects. Under this model framework, the working covariance matrix can be constructed easily for parameter estimation using the generalized estimating equation approach.
The generalized estimating equations approach is widely used in analyzing
correlated data with marginal models. But there is little discussion about the procedure of model fitting when using this approach. We consider using some existing nonparametric/semiparametric methods for checking the functional form of continuous covariate. To illustrate the proposed model and the fitting strategy, we analyze metabolic disorder data from 2588 families included in three survey cycles of a community-based cardiovascular disease study in Chu-Dong and Pu-Tzu.
|
|---|
